1 /*
2 * Copyright (c) 2018-2020 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24 #include "src/core/CL/kernels/CLComparisonKernel.h"
25
26 #include "arm_compute/core/CL/CLHelpers.h"
27 #include "arm_compute/core/CL/ICLTensor.h"
28 #include "src/core/CL/CLValidate.h"
29 #include "src/core/helpers/AutoConfiguration.h"
30 #include "src/core/helpers/WindowHelpers.h"
31 #include "support/StringSupport.h"
32
33 #include <map>
34
35 namespace arm_compute
36 {
37 namespace
38 {
39 // Create supported comparisons map
40 const std::map<ComparisonOperation, std::string> supported_comparison_ops =
41 {
42 { ComparisonOperation::Equal, "EQUAL" },
43 { ComparisonOperation::NotEqual, "NOTEQUAL" },
44 { ComparisonOperation::Greater, "GREATER" },
45 { ComparisonOperation::GreaterEqual, "GREATEREQUAL" },
46 { ComparisonOperation::Less, "LESS" },
47 { ComparisonOperation::LessEqual, "LESSEQUAL" },
48 };
49
calculate_num_elems_processed_per_iteration(const ITensorInfo & input)50 int calculate_num_elems_processed_per_iteration(const ITensorInfo &input)
51 {
52 return 16 / input.element_size();
53 }
54
validate_arguments(const ITensorInfo & input1,const ITensorInfo & input2,const ITensorInfo & output,ComparisonOperation operation)55 Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output, ComparisonOperation operation)
56 {
57 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&input1);
58 ARM_COMPUTE_RETURN_ERROR_ON(input1.data_type() == DataType::UNKNOWN);
59 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input1, &input2);
60 ARM_COMPUTE_RETURN_ERROR_ON(supported_comparison_ops.count(operation) == 0);
61
62 const TensorShape out_shape = TensorShape::broadcast_shape(input1.tensor_shape(), input2.tensor_shape());
63 ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible");
64
65 // Validate in case of configured output
66 if(output.total_size() > 0)
67 {
68 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&output, 1, DataType::U8);
69 ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output.tensor_shape(), 0),
70 "Wrong shape for output");
71 }
72
73 return Status{};
74 }
75
validate_and_configure_window(ITensorInfo & input1,ITensorInfo & input2,ITensorInfo & output)76 std::pair<Status, Window> validate_and_configure_window(ITensorInfo &input1, ITensorInfo &input2, ITensorInfo &output)
77 {
78 const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(input1, input2);
79 const TensorShape &out_shape = broadcast_pair.first;
80 const ValidRegion &valid_region = broadcast_pair.second;
81
82 const unsigned int num_elems_processed_per_iteration = calculate_num_elems_processed_per_iteration(input1);
83
84 // Auto initialize output if not initialized
85 auto_init_if_empty(output, out_shape, 1, DataType::U8, QuantizationInfo());
86
87 Window win = calculate_max_window(valid_region, Steps(num_elems_processed_per_iteration));
88 Window win_input1 = win.broadcast_if_dimension_le_one(input1);
89 Window win_input2 = win.broadcast_if_dimension_le_one(input2);
90
91 AccessWindowHorizontal input1_access(&input1, 0, num_elems_processed_per_iteration);
92 AccessWindowHorizontal input2_access(&input2, 0, num_elems_processed_per_iteration);
93 AccessWindowHorizontal output_access(&output, 0, num_elems_processed_per_iteration);
94
95 bool window_changed = update_window_and_padding(win_input1, input1_access)
96 || update_window_and_padding(win_input2, input2_access)
97 || update_window_and_padding(win, output_access);
98
99 output_access.set_valid_region(win, valid_region);
100
101 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
102 return std::make_pair(err, win);
103 }
104 } // namespace
105
CLComparisonKernel()106 CLComparisonKernel::CLComparisonKernel()
107 : _input1(nullptr), _input2(nullptr), _output(nullptr)
108 {
109 }
110
configure(const ICLTensor * input1,const ICLTensor * input2,ICLTensor * output,ComparisonOperation operation)111 void CLComparisonKernel::configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, ComparisonOperation operation)
112 {
113 configure(CLKernelLibrary::get().get_compile_context(), input1, input2, output, operation);
114 }
115
configure(const CLCompileContext & compile_context,const ICLTensor * input1,const ICLTensor * input2,ICLTensor * output,ComparisonOperation operation)116 void CLComparisonKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, ComparisonOperation operation)
117 {
118 ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
119 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*input1->info(), *input2->info(), *output->info(), operation));
120
121 // Configure kernel window
122 auto win_config = validate_and_configure_window(*input1->info(), *input2->info(), *output->info());
123 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
124
125 _input1 = input1;
126 _input2 = input2;
127 _output = output;
128
129 const std::string &operation_name = supported_comparison_ops.at(operation);
130 std::string kernel_name = "compare_" + lower_string(operation_name);
131
132 // Set kernel build options
133 std::set<std::string> build_opts;
134 build_opts.emplace("-DDATA_TYPE=" + get_cl_type_from_data_type(input1->info()->data_type()));
135 build_opts.emplace("-DVEC_SIZE=" + support::cpp11::to_string(calculate_num_elems_processed_per_iteration(*input1->info())));
136 build_opts.emplace("-DOP=" + operation_name);
137 build_opts.emplace("-DOP_NAME=" + lower_string(operation_name));
138 if(is_data_type_quantized(input1->info()->data_type()))
139 {
140 const UniformQuantizationInfo iq1_info = input1->info()->quantization_info().uniform();
141 const UniformQuantizationInfo iq2_info = input2->info()->quantization_info().uniform();
142
143 build_opts.emplace("-DOFFSET_IN1=" + support::cpp11::to_string(iq1_info.offset));
144 build_opts.emplace("-DOFFSET_IN2=" + support::cpp11::to_string(iq2_info.offset));
145 build_opts.emplace("-DSCALE_IN1=" + float_to_string_with_full_precision(iq1_info.scale));
146 build_opts.emplace("-DSCALE_IN2=" + float_to_string_with_full_precision(iq2_info.scale));
147 kernel_name += "_quantized";
148 }
149
150 // Create kernel
151 _kernel = create_kernel(compile_context, kernel_name, build_opts);
152
153 ICLKernel::configure_internal(win_config.second);
154
155 // Set config_id for enabling LWS tuning
156 _config_id = kernel_name;
157 _config_id += "_";
158 _config_id += lower_string(string_from_data_type(input1->info()->data_type()));
159 _config_id += "_";
160 _config_id += support::cpp11::to_string(output->info()->dimension(0));
161 _config_id += "_";
162 _config_id += support::cpp11::to_string(output->info()->dimension(1));
163 _config_id += lower_string(string_from_data_layout(input1->info()->data_layout()));
164 }
165
validate(const ITensorInfo * input1,const ITensorInfo * input2,const ITensorInfo * output,ComparisonOperation operation)166 Status CLComparisonKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, ComparisonOperation operation)
167 {
168 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output);
169
170 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*input1, *input2, *output, operation));
171 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(*input1->clone(), *input2->clone(), *output->clone()).first);
172
173 return Status{};
174 }
175
run(const Window & window,cl::CommandQueue & queue)176 void CLComparisonKernel::run(const Window &window, cl::CommandQueue &queue)
177 {
178 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
179 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
180
181 const TensorShape &in_shape1 = _input1->info()->tensor_shape();
182 const TensorShape &in_shape2 = _input2->info()->tensor_shape();
183 const TensorShape &out_shape = _output->info()->tensor_shape();
184
185 bool can_collapse = true;
186 const bool is_vector = in_shape1.num_dimensions() == 1 || in_shape2.num_dimensions() == 1;
187 if(std::min(in_shape1.total_size(), in_shape2.total_size()) > 1 && !is_vector)
188 {
189 can_collapse = (std::min(in_shape1.num_dimensions(), in_shape2.num_dimensions()) > Window::DimZ);
190 for(size_t d = Window::DimZ; can_collapse && (d < out_shape.num_dimensions()); d++)
191 {
192 can_collapse = (in_shape1[d] == in_shape2[d]);
193 }
194 }
195
196 bool has_collapsed = false;
197 Window collapsed = can_collapse ? window.collapse_if_possible(ICLKernel::window(), Window::DimZ, &has_collapsed) : window;
198
199 const TensorShape &in_shape1_collapsed = has_collapsed ? in_shape1.collapsed_from(Window::DimZ) : in_shape1;
200 const TensorShape &in_shape2_collapsed = has_collapsed ? in_shape2.collapsed_from(Window::DimZ) : in_shape2;
201
202 Window slice = collapsed.first_slice_window_3D();
203 Window slice_input1 = slice.broadcast_if_dimension_le_one(in_shape1_collapsed);
204 Window slice_input2 = slice.broadcast_if_dimension_le_one(in_shape2_collapsed);
205
206 do
207 {
208 unsigned int idx = 0;
209
210 add_3D_tensor_argument(idx, _input1, slice_input1);
211 add_3D_tensor_argument(idx, _input2, slice_input2);
212 add_3D_tensor_argument(idx, _output, slice);
213
214 enqueue(queue, *this, slice, lws_hint());
215
216 ARM_COMPUTE_UNUSED(collapsed.slide_window_slice_3D(slice_input1));
217 ARM_COMPUTE_UNUSED(collapsed.slide_window_slice_3D(slice_input2));
218 }
219 while(collapsed.slide_window_slice_3D(slice));
220 }
221
border_size() const222 BorderSize CLComparisonKernel::border_size() const
223 {
224 const int num_elems_processed_per_iteration = calculate_num_elems_processed_per_iteration(*_input1->info());
225
226 const unsigned int replicateSize = _output->info()->dimension(0) - std::min(_input1->info()->dimension(0), _input2->info()->dimension(0));
227 const unsigned int border = std::min<unsigned int>(num_elems_processed_per_iteration - 1U, replicateSize);
228 return BorderSize{ 0, border, 0, 0 };
229 }
230 } // namespace arm_compute
231